Abstract: According to the problems in the fault diagnosis method of power transformer, such as incomplete code, less accurate rate and so on, this paper studied the neural network algorithm of Bayesian theory, and proposed a transformer oil chromatographic fault diagnosis method which was based on the L-M (Levenberg-Marquardt) neural network optimized by Bayesian regularization algorithm. Firstly, the algorithm used Bayesian approach to determine the hyper parameters, which could make the neural network adaptively adjust the parameter in the training process and get the optimization parameters of the objective function. Secondly, the method used the fuzzy theory to handle the boundary of improved three-ratio method, and then the characteristic gas ratio code was obtained and used as the network model input, which had advantages to remove the redundant information and overcome the absolute of code boundary. Finally, this paper used the simulation software to simulate the operation data of typical transformer and verified the feasibility of the proposed algorithm. The results show that the iteration times is 21, and the error square sum of actual and predicted values is only 0.000 618, when the proposed model is used for the fault diagnosis of transformer

Key words: power transformer, Bayesian regularization, hyper parameters, neural network, fuzzy theory